topic Re: Buffer hits and kernel parms in Operating System - HP-UX

Buffer hits and kernel parms

Sean OB_1 — Mon, 28 Oct 2002 15:59:52 GMT

Hello,

I have a L2000 that we just increaded memory from 7GB to 16GB. Accordingly I modified the kernel parms:

dbc_min_pct=2
dbc_max_pct=8

Which should give me over 1GB of memory for buffers.

However my buffer hits have dropped from near 100% to the following:

09:56:11 bread/s lread/s %rcache bwrit/s lwrit/s %wcache pread/s pwrit/s
09:56:13 640 1921 67 11 35 69 0 0
09:56:15 708 1971 64 6 22 70 0 0
09:56:17 757 2228 66 31 32 2 0 0
09:56:19 639 1646 61 8 22 60 0 0
09:56:21 646 1427 55 9 28 68 0 0
09:56:23 602 1464 59 50 18 0 0 0
09:56:25 719 1236 42 14 55 75 0 0
09:56:27 730 1377 47 8 8 0 0 0
09:56:29 696 1582 56 9 17 46 0 0
09:56:31 693 2020 66 14 29 53 0 0

Average 683 1687 60 16 27 40 0 0

Are there any other kernel parms I should look at? Should I increase the max %? I would think 1GB would be more than adequate.

TIA,

Sean

Re: Buffer hits and kernel parms

harry d brown jr — Mon, 28 Oct 2002 16:05:49 GMT

What kind of data is being read/written? Database?

Actually 8% max of 16GB is 1.2GB is usually rather high - but it depends again on what kind of IO you are doing. I like to keep mine under 400MB.

live free or die
harry

Re: Buffer hits and kernel parms

A. Clay Stephenson — Mon, 28 Oct 2002 16:09:05 GMT

Your dynamic buffer cache could be overriden by a non-zero bufpages setting but your main problem is that your sampling period is much too short for these measurements. You need to sample over a much longer period of time and while the box is under different loads before any kind of intelligent estimate can be made as to the 'correct' values for buffer cache.

I doubt that you are going to see any real improvement over about 800-1GB in 11.11 and at most about 400-500 MB in 11.0.

Re: Buffer hits and kernel parms

Sean OB_1 — Mon, 28 Oct 2002 16:26:01 GMT

I've actually been sampling for quite a while at various times during the day and this is about the average that I've seen.

The I/O is for a non relational database (Unidata) going to an FC60 array.

bufpages is set to zero.

I'll sample for the entire day today and see if the results are any different.